The rapid advancement of artificial intelligence has fueled expectations that autonomous AI agents will transform businesses, increase productivity, and eventually perform complex economic tasks with minimal human intervention.
However, a new benchmark released by SemiAnalysis challenges these expectations, revealing that the leading model scored just 12% on economically viable tasks evaluated through its GDPx agentic benchmark.
The findings highlight a significant gap between AI’s impressive demonstrations and its ability to deliver reliable economic value in real-world environments.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Agentic AI refers to systems capable of planning, executing multistep workflows, using digital tools, and making decisions toward specific objectives. Unlike conventional chatbots that primarily respond to prompts, AI agents are designed to complete assignments with limited human supervision.
These capabilities have attracted substantial investment from technology companies seeking to automate administrative work, software development, financial analysis, customer service, and other business operations.
The GDPx benchmark suggests that successfully completing economically valuable tasks remains a considerable challenge. A score of 12% for the top-performing model indicates that most evaluated tasks did not meet the benchmark’s requirements for economic viability.
Although this result does not mean AI systems are incapable of performing useful work, it raises questions about how quickly autonomous agents can become dependable substitutes for human labor.
One important distinction is that producing a convincing response is not equivalent to completing a business assignment. An agent might generate a report, write code, or propose a strategy, but genuine economic usefulness requires accuracy, consistency, and successful execution.
In professional environments, even small mistakes can generate additional costs, expose organizations to operational risks, or require extensive human intervention. The findings emphasize the importance of evaluating AI beyond traditional academic examinations and coding competitions.
Models can perform impressively on standardized tests while struggling with tasks involving ambiguous instructions, changing circumstances, unfamiliar software, and multiple interconnected decisions. Real businesses operate under these conditions daily, making practical reliability essential to successful automation.
For companies investing heavily in artificial intelligence, the implications are significant. Organizations must distinguish between tasks that AI can already perform effectively and workflows that still require substantial human oversight.
Rather than replacing entire departments immediately, businesses may achieve better results by deploying AI agents for clearly defined activities, measuring their performance, and expanding their responsibilities as reliability improves.
The benchmark could influence competition among AI developers. Companies such as OpenAI, Anthropic, and Google are investing in increasingly capable models and agentic systems.
However, the ability to translate technical sophistication into measurable economic productivity may become a more important competitive advantage than impressive demonstrations or benchmark scores alone.
The 12% result should not be interpreted as a permanent ceiling for agentic AI. Performance can improve through better models, stronger reasoning capabilities, improved tool integration, and more effective verification systems.
Carefully designed workflows can also reduce the likelihood of errors and help agents operate within clearly established boundaries. SemiAnalysis’s GDPx benchmark introduces an important reality check into the artificial intelligence debate.
The technology has considerable potential, but widespread economic transformation depends on more than intelligence alone. AI agents must demonstrate that they can complete valuable work accurately, consistently, and affordably.
Until that happens at a much larger scale, businesses should approach autonomous AI with ambition balanced by evidence. The next stage of the AI revolution will depend not simply on what models can demonstrate, but on how much useful work they can reliably deliver.